Modelpublic

davidramos/int8-whisper-base-v2-asr-bengali

sourceHugging Faceupdated 3y agoView on Hugging Face

0likes16downloads

Model Card

Training procedure

The following bitsandbytes quantization config was used during training:

—quant_method: bitsandbytes
—loadin8bit: True
—loadin4bit: False
—llmint8threshold: 6.0
—llmint8skip_modules: None
—llmint8enablefp32cpu_offload: False
—llmint8hasfp16weight: False
—bnb4bitquant_type: fp4
—bnb4bitusedoublequant: False
—bnb4bitcompute_dtype: float32

Framework versions

—PEFT 0.5.0

Training loss

Epoch | Training Loss | Validation Loss 1 | 0.573400 | 0.595536 2 | 0.476500 | 0.506768 3 | 0.421000 | 0.472346